AI in Healthcare

The framework was preliminarily evaluated on a publicly available single-cell time-lapse microscopy dataset, which, while not medical, shares similar temporal dynamics and varying external conditions (glucose-lactose) akin to EHR parameters. This dataset allows for validation of core modeling principles: learning trajectories, integrating time-varying conditions, and forecasting future states. The ConvLSTM model, specifically using the ST-FiLM algorithm for spatio-temporal condition embedding, was implemented. Three models were compared: a long-term prediction model (Model 1) and two short-term variants (Models 2 and 3).

Model 1, the long-term ConvLSTM model, demonstrated superior performance with the minimum test loss value and maximum SSIM. Visually, Model 1's predictions more closely resembled ground truth images, retaining fine structural details that short-term models lost over time. This confirms the efficacy of incorporating longer historical sequences and condition embeddings for accurate spatio-temporal forecasting.

Future efforts will focus on implementing and validating the proposed ConvLSTM- and ViViT-based pipelines on real patient datasets. Identified datasets include OASIS-3, ADNI, and NACC for Alzheimer's disease (longitudinal MRI with EHRs), and KIOS for mammography. Image registration will be a critical preprocessing step for medical datasets, with modality-specific SSIM and mutual information (MI) thresholds established to ensure robust performance. Initial focus will be on 2D mammogram data with EHRs, expanding to 3D MRI data by utilizing reduced regions of interest (ROIs).